Kelvin Lee
@inequivariant.bsky.social
Scientific ML Engineer @ NVIDIA working on AI4Science. Previously postdoc at MIT Chemistry and Smithsonian Astrophysical Observatory.
Opinions are my own 😽
Opinions are my own 😽
I think your hartree to meV is off by a 1000 (it should be around 27 eV)
January 21, 2025 at 2:30 PM
I think your hartree to meV is off by a 1000 (it should be around 27 eV)
We hope that the work will spark some discussion throughout the MLIP community - users and developers alike - to pay attention not just to benchmark metrics, but also qualitative assessments of what their models are learning.
Slides are available here: laserkelvin.github.io/neurips24-ai...
Slides are available here: laserkelvin.github.io/neurips24-ai...
Deconstructing equivariant representations in molecular systems
laserkelvin.github.io
December 12, 2024 at 3:15 PM
We hope that the work will spark some discussion throughout the MLIP community - users and developers alike - to pay attention not just to benchmark metrics, but also qualitative assessments of what their models are learning.
Slides are available here: laserkelvin.github.io/neurips24-ai...
Slides are available here: laserkelvin.github.io/neurips24-ai...
The spotlight talk will be on our recent work diving into what gets learned by #equivariant neural networks used for interatomic potentials.
The key finding was that higher order (non-scalar) latent spaces were largely being ignored, which leaves modeling performance on the table.
The key finding was that higher order (non-scalar) latent spaces were largely being ignored, which leaves modeling performance on the table.
December 12, 2024 at 3:14 PM
The spotlight talk will be on our recent work diving into what gets learned by #equivariant neural networks used for interatomic potentials.
The key finding was that higher order (non-scalar) latent spaces were largely being ignored, which leaves modeling performance on the table.
The key finding was that higher order (non-scalar) latent spaces were largely being ignored, which leaves modeling performance on the table.